Evaluating the accuracy of satellite-based methods to estimate residential proximity to agricultural crops

Epidemiologic investigations increasingly employ remote sensing data to estimate residential proximity to agriculture as a means of approximating individual-level pesticide exposure. Few studies have examined the accuracy of these methods and the implications for exposure misclassification.

Compare metrics of residential proximity to agricultural land between a groundtruth approach and commonly-used satellite-based estimates.

We inspected 349 fields and identified crops in current production within a 0.5 km radius of 40 residences in Idaho. We calculated the distance from each home to the nearest agricultural field and the total acreage of agricultural fields within a 0.5 km buffer. We compared these groundtruth estimates to satellite-derived estimates from three widely used datasets: CropScape, the National Land Cover Database (NLCD), and Landsat imagery (using Normalized Difference Vegetation Index thresholds).

We found poor to moderate agreement between the classification of individuals living within 0.5 km of an agricultural field between the groundtruth method and the comparison datasets (53.1-77.6%). All satellite-derived estimates overestimated the acreage of agricultural land within 0.5 km of each home (average = 82.8-148.9%). Using two satellite-derived datasets in conjunction resulted in substantial improvements; specifically, combining CropScape or NLCD with Landsat imagery had the highest percent agreement with the groundtruth data (92.8-93.8% agreement).

Residential proximity to agriculture is frequently used as a proxy for pesticide exposure in epidemiologic investigations, and remote sensing-derived datasets are often the only practical means of identifying cultivated land. We found that estimates of agricultural proximity obtained from commonly-used satellite-based datasets are likely to result in exposure misclassification. We propose a novel approach that capitalizes on the complementary strengths of different sources of satellite imagery, and suggest the combined use of one dataset with high temporal resolution (e.g., Landsat imagery) in conjunction with a second dataset that delineates agricultural land use (e.g., CropScape or NLCD).